Adaptive transformations in model driven machine learning

Speaker:

Tamás Dózsa, Eötvös Loránd University, Budapest

Abstract:

The ability of machine learning (ML) methods to approximate highly nonlinear operators allowed for remarkable advancments in signal and image processing. Current state-of-the-art deep learning models however are often described by millions of parameters. This makes the interpretability of trained models difficult and their deployment costly. In the presentation, so-called model driven machine learning methods are discussed which allow for simpler and more interpretable model architectures. Model driven methods replace certain parts of ML models with adaptive mathematical transformations. If the transformations are chosen in an appropriate manner their parameters will carry exact physical meaning. A general framework to construct model driven methods is given utlizing variable projection operators. Several concrete model driven ML approaches are reviewed using wavelet transformations and adaptive orthogonal expansions. Applications in fault detection systems and biomedical signal processing are discussed.
 

Biography:

Tamás Dózsa is a fourth year PhD student from the Numerical Analysis department of Eötvös Loránd University Budapest, Hungary. He is expected to defend his dissertation in the autumn of 2024. He also works as a researcher in the Systems and Control Laboratory of the HUN-REN Institute for Computer Science and Control (SZTAKI). His research focuses on the study of various adaptive orthogonal transformations and their application in control theory, machine learning and signal processing. He has written 20 scientific papers thus far including works published in 1 Q1 and 3 D1 ranked international journals. He has been the recepient of several scholarships and awards including the Cooperative Doctoral Programme scholarship by the Hungarian Ministry of Innovation and Technology and several outstanding PhD student awards from HUN-REN SZTAKI.

Host: Prof. R. Krause